A Method for Enhancing Conversational Recommendation Systems by Utilizing Adjacent Information

Abstract

A method for enhancing a conversational recommendation system using adjacent information, which belongs to the field of computer artificial intelligence technology. The present invention solves the problems of sparse user preferences and insufficient training data existing in the existing conversational recommendation system. The present invention retrieves adjacent items and uses the current conversation and adjacent items to construct pseudo-label data to alleviate the problem of insufficient training set data and enhance the training process of the conversational recommendation system; retrieves adjacent conversations and uses the information of adjacent conversations to alleviate the challenge of sparse user preferences and enhance the process of predicting items in the conversational recommendation system; the method of the present invention plays a very good role in alleviating the challenges of sparse user preferences and insufficient training data in the conversational recommendation system, and can significantly improve the item recommendation and response generation effects of the conversational recommendation system. The method of the present invention can be applied to the field of computer artificial intelligence.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer artificial intelligence, and particularly relates to a method for enhancing a conversational recommendation system by using adjacent information. Background Art

[0002] The conversational recommendation system is an important research task in information retrieval in artificial intelligence technology. The recommendation system aims to predict the user's preference for items based on the historical interaction information between the user and the items. The recommendation system can select the information that the user may be interested in from a large amount of information and recommend it to the user, enabling the rapid rise of platforms such as e-commerce and short videos. However, the recommendation system cannot obtain the user's real-time information needs and preferences, and the recommendations provided by the system have certain limitations.

[0003] The conversational recommendation system aims to interact with the user through natural language, obtain the user's real-time preferences and feedback during the chat process, and provide appropriate recommendation results for the user. In addition, compared with the traditional recommendation system, the conversational recommendation system can also integrate the recommendation results into the system response to form a recommendation reason, making it easier for the user to accept the recommendation.

[0004] The conversational recommendation system has two challenges: (1) Modeling sparse preferences in the conversation (2) Lack of sufficient training data.

[0005] Sparse user preferences. The interaction in natural language enables the conversational recommendation system to obtain a wider range of user preferences, but also brings great challenges to understanding user preferences. In traditional recommendation systems, the system understands user preferences based on a certain number of clear and specific user behaviors, such as a user clicking on a product or purchasing a product. In the conversational recommendation scenario, however, the needs and preferences expressed by users in a conversation are relatively few. In an 8-round conversation, perhaps only 1 round has the user expressing useful needs and preferences, such as "I've watched Doctor Strange, it's very interesting, and I also want to watch some other movies with superheroes". Therefore, a great deal of work has explored using external resources to enhance the understanding of the items involved in the conversation. Chen et al. introduced DBpedia entities to enable the conversational recommendation system to understand movie entities involved in the conversation, such as "Doctor Strange" (Chen Q, Lin J, Zhang Y, et al. Towards knowledge-based recommender dialog system [A]. EMNLP-IJCNLP 2019 - 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint Conference on Natural Language Processing, Proceedings of the Conference [C]. 2019:1803–1813.). Zhou et al. introduced ConceptNet entities to enable the conversational recommendation system to understand "superheroes" and constructed the association between "superheroes" and "Doctor Strange" (Zhou K, Zhao W X, Bian S, et al. Improving Conversational Recommender Systems via Knowledge Graph based Semantic Fusion [A]. Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining [C]. 2020, 20:1006–1014.).Lu et al. further introduced movie-related reviews and used the sentiment in the reviews to train the conversational recommendation system to understand the user's sentiment towards movie entities (Lu Y, Bao J, Song Y, et al. RevCore: Review-augmented Conversational Recommendation [A]. Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 [C]. 2021: 1161–1173.). However, the challenge of sparse user preferences still limits the further improvement of the performance of the conversational recommendation system.

[0006] Insufficient training data. The conversational recommendation system needs to learn the association between the given dialogue context and the recommended items from the training data. Due to the challenge of sparse user preferences in the conversational recommendation scenario, the conversational recommendation system requires a large amount of training data to capture the association between sparse preferences and target items. However, the datasets available for existing conversational recommendation systems are usually small in scale, and constructing a large-scale and high-quality dataset required for building a conversational recommendation system requires significant human and time resources (Li R, Kahou S E, Schulz H, et al. Towards deep conversational recommendations [J]. Advances in Neural Information Processing Systems, 2018, 2018-Decem (NeurIPS): 9725–9735.).

[0007] In summary, the existing conversational recommendation systems mainly have the following problems:

[0008] I. Sparse user preferences: In the conversational recommendation scenario, the user's expressed needs and preferences in a dialogue are few.

[0009] II. Insufficient training data: The datasets available for existing conversational recommendation systems are small in scale. Summary of the Invention

[0010] The purpose of the present invention is to solve the problems of sparse user preferences and insufficient training data existing in the existing conversational recommendation systems, and to propose a method for enhancing a conversational recommendation system using adjacent information.

[0011] The technical solution adopted by the present invention to solve the above technical problems is:

[0012] A method for enhancing a conversational recommendation system using adjacent information, the method specifically includes the following steps:

[0013] Step 1: Construct pseudo-label data using adjacent items of the target item recommended in the conversation, and then construct a training set for the conversational recommendation system based on the pseudo-label data. After training, obtain a trained conversational recommendation system;

[0014] The specific process of Step 1 is as follows:

[0015] Step 1: Obtain a conversation dataset;

[0016] Step 2: For any conversation in the dataset, construct a training sample using this conversation where, represents the representation vector of the conversation context obtained through the conversational recommendation system, represents the representation vector of the target item recommended in this conversation, represents the representation vector of the reply generated for this conversation;

[0017] Using the representation vector of the target item Retrieve n adjacent items of the target item from the set of candidate items Y to be recommended. Use the representation vectors of the retrieved adjacent items to replace the training sample in Take the replaced training sample as a pseudo-label training sample;

[0018] Similarly, process each conversation in the dataset separately. Denote the set composed of the training samples constructed based on each conversation in the dataset as D, and denote the set of all obtained pseudo-label training samples as where, represents the set of adjacent items retrieved according to the representation vector of a target item, represents the set the number of adjacent items contained in, that is |D| represents the number of conversations contained in the dataset, and i' represents the i'-th pseudo-label training sample in D p in, represents the representation vector of the conversation context corresponding to the i'-th pseudo-label training sample, represents the representation vector of the recommended target item in the i'-th pseudo-label training sample, represents the representation vector of the reply generated for the conversation corresponding to the i'-th pseudo-label training sample;

[0019] Step 3: Take the union D∪D p of D and D p as the training set for this iteration, and use the training set to train the conversational recommendation system

[0020] Step 4: Repeat the process of Step 2 to Step 3 until the iteration stops when the set maximum number of iterations is reached, and a trained conversational recommendation system is obtained.

[0021] Step 2: Input the current conversation into the trained conversational recommendation system to obtain the item recommendation result and reply generation result for the current conversation.

[0022] The beneficial effects of the present invention are:

[0023] The present invention retrieves adjacent items and uses the current conversation and adjacent items to construct pseudo-label data to alleviate the problem of insufficient training set data and enhance the training process of the conversational recommendation system; retrieves adjacent conversations and uses the information of adjacent conversations to alleviate the challenge of sparse user preferences and enhance the process of predicting items by the conversational recommendation system; the method of the present invention plays a very good role in alleviating the challenges of sparse user preferences and insufficient training data of the conversational recommendation system, and can significantly improve the item recommendation and reply generation effects of the conversational recommendation system. Description of the Drawings

[0024] Figure 1 It is a flowchart of recommending items for a traditional conversational recommendation system;

[0025] Figure 2 It is a flowchart of recommending items by enhancing a conversational recommendation system using adjacent information according to the present invention. Detailed Embodiments

[0026] Detailed Embodiment 1: Combine Figure 2 to illustrate this embodiment. The method for enhancing a conversational recommendation system using adjacent information described in this embodiment specifically includes the following steps:

[0027] Step 1: Use the adjacent items of the target item to be recommended in the conversation to construct pseudo-label data, and then construct a training set for the conversational recommendation system based on the pseudo-label data. After training, a trained conversational recommendation system is obtained;

[0028] The specific process of Step 1 is as follows:

[0029] Step 1: Obtain a conversation dataset;

[0030] Step 2: For any conversation in the dataset, use the conversation to construct a training sample where, represents the representation vector of the conversation context obtained through the conversational recommendation system, represents the representation vector of the target item to be recommended in the conversation, represents the representation vector of the reply generated for the conversation;

[0031] Using the representation vector of the target item Retrieve n items adjacent to the target item from the set Y of candidate items to be recommended, and use the representation vectors of the retrieved adjacent items to replace the training samples in The replaced training samples are used as pseudo-label training samples;

[0032] Similarly, each conversation in the dataset is processed separately. Denote the set composed of the training samples constructed based on each conversation in the dataset as D, and denote the set composed of all the obtained pseudo-label training samples as wherein, represents the set of adjacent items retrieved according to the representation vector of a target item, represents the set the number of adjacent items contained in the set, that is |D| represents the number of conversations contained in the dataset, and i' represents the i'-th pseudo-label training sample in D p in, represents the representation vector of the conversation context corresponding to the i'-th pseudo-label training sample, represents the representation vector of the recommended target item in the i'-th pseudo-label training sample, represents the representation vector for generating a response to the conversation corresponding to the i'-th pseudo-label training sample;

[0033] Step 3: Take the union D∪D p of D and D p as the training set for this iteration, and use the training set to train the conversational recommendation system

[0034] Step 4: Repeat the process of Step 2 to Step 3 until the set maximum number of iterations is reached, and then stop the iteration to obtain the trained conversational recommendation system

[0035] Step 2: Input the current conversation into the trained conversational recommendation system to obtain the item recommendation result and response generation result for the current conversation.

[0036] In Step 1, the vectors used in the current iteration are all obtained based on the conversational recommendation system obtained in the previous iteration.

[0037] As Figure 1 shown, it is the flowchart of item recommendation by a traditional conversational recommendation system. Compared with the traditional system, the method of the present invention can significantly improve the system in both item recommendation and response generation effects.

[0038] Specific Embodiment 2: This embodiment is a further limitation of Specific Embodiment 1. The representation vector of the target item is used to retrieve n items adjacent to the target item from the set Y of candidate items to be recommended. The specific process is as follows:

[0039] Calculate the cosine similarity between the representation vector of each item in the set Y of candidate items to be recommended and respectively, and use the n items with the largest cosine similarity to as the n retrieved adjacent items:

[0040]

[0041] Among them, the function represents retrieving n adjacent items from the set Y of candidate items to be recommended.

[0042] Specific Embodiment 3: This embodiment is a further limitation of Specific Embodiment 2. The specific process of step 3 is as follows:

[0043] Step 1): For any training sample in the union D∪D p (this training sample can be a training sample in D or a pseudo-label training sample in D p ), retrieve the dialogue representation vector adjacent to the context representation vector of this training sample from the set X of candidate dialogue representation vectors, and calculate the parameter b of each candidate item based on the retrieved adjacent dialogue representation vector, as well as the vector ′ corresponding to this training sample

[0044] According to the vector and the parameter b ′ to obtain the item recommendation result and reply generation result for the context representation vector ;

[0045] Similarly, process each training sample in the union D∪D p ;

[0046] Step 2): Calculate the loss of the item recommendation task according to the item recommendation result and reply generation result obtained in step 1), and then update the parameters of the dialogue recommendation system ;

[0047] Step 3): Set the pseudo-label training sample set to an empty set, that is, set and then return to step 2.

[0048] Specific Embodiment 4: This embodiment is a further limitation of Specific Embodiment 3. The vector x′ The calculation process is as follows:

[0049] Given the set X of candidate dialogue context representation vectors, calculate the cosine similarity between each vector in X and the vector respectively. Take the m dialogues corresponding to the largest m cosine similarities as the adjacent dialogues of the vector ;

[0050]

[0051] Among them, represents the set composed of adjacent dialogues, and φ(x, X, m) represents retrieving m dialogues adjacent to the vector from the set X;

[0052] Then, the present invention uses to calculate the representation vector

[0053]

[0054] Among them, v i is the intermediate vector of the i-th adjacent dialogue, w i is the weight of the i-th adjacent dialogue, represents the number of adjacent dialogues in the set , R represents real numbers, and d is the dimension of .

[0055] Specific Embodiment 5: This embodiment further limits Specific Embodiment 4. The v i is calculated according to the representation vector of the i-th adjacent dialogue and the representation vector of the recommended item of the i-th adjacent dialogue:

[0056]

[0057] Among them, MLP is a multi-layer perceptron.

[0058] Specific Embodiment 6: This embodiment further limits Specific Embodiment 5. The w i is calculated according to the representation vector of the i-th adjacent dialogue and :

[0059]

[0060] Among them, Linear is a linear layer, and Sigmoid is an activation function.

[0061] Embodiment Seven: This embodiment further limits Embodiment Six, and the MLP consists of two linear layers.

[0062] Embodiment Eight: This embodiment further limits Embodiment Seven, and the calculation process of the parameter b ′ is as follows:

[0063] b ′ = b + b s ,

[0064] where b and b s are intermediate parameters;

[0065] Embodiment Nine: This embodiment further limits Embodiment Eight, and the calculation process of the intermediate parameters b and b s is as follows:

[0066] For the candidate item e, the range of e is the item set E:

[0067]

[0068] where τ k represents the number of times the candidate item e appears in the k-th conversation, w k represents the weight of the k-th conversation, and the weight of the current conversation is 1;

[0069]

[0070] where, based on b, the present invention introduces other adjacent items through item-item cosine similarity, T k,j represents the number of times the j-th item appears in the k-th conversation, E is the item set, |E| is the number of items in the item set, and s j is the cosine similarity between the representation vector of the j-th item in the item set and the representation vector of the candidate item e;

[0071] Adding b and b s gives the parameter b corresponding to the candidate item e ′ ;

[0072] Similarly, the parameter b corresponding to each item in the item set E is obtained ′ .

[0073] Embodiment Ten: This embodiment further limits Embodiment Nine, and the specific process of Step 2 is as follows:

[0074] Step ①: Calculate the probability of being recommended for each candidate item respectively:

[0075] The conversational recommendation system involves an item recommendation task and a response generation task. The item recommendation task is defined as the system recommending target items given the conversation context. The conversational recommendation system usually uses the representation vector of the conversation context the representation vector of candidate items and a learnable parameter b to calculate the probability of recommending item y.

[0076]

[0077] where, represents the context representation vector of the current conversation, the superscript T represents transpose, represents the representation vector of candidate items, represents the probability that the item corresponding to the vector is recommended; b ′ is the b ′ parameter value corresponding to the item corresponding to, here is calculated according to ;

[0078] The candidate item with the highest recommended probability is taken as the item recommendation result;

[0079] The vectors used in step ① are all obtained based on the trained conversational recommendation system;

[0080] Step ②, after obtaining the recommended item, retrieve the adjacent items of the recommended item and generate a response according to the retrieved adjacent items:

[0081]

[0082] where, Decoder is the decoder, X 0 is the context of the current conversation, is the representation vector of the recommended item, select Q items (the value of Q in the present invention is 10) with the largest b' parameter values from the retrieved adjacent items, fuse the representation vectors of the selected items to obtain the fusion result

[0083] is the first,..., the (q - 1)-th word vector in the generated response, represents the probability of being the q-th word vector in the generated response;

[0084] Take the word vector corresponding to the maximum probability as the q-th word vector of the generated response.

[0085] Generate each word in the response in turn using the method of step ②, and finally obtain the generated response. When generating the response, the dialogue recommendation system decodes word by word, that is, given the first q - 1 word vectors The representation vector of the current dialogue context And the representation vector of the item recommended by the dialogue recommendation system Predict the probability of the q-th word. At the same time, the present invention introduces the fusion of the representation vectors of multiple adjacent items, that is The present invention utilizes Replace Enable the system to utilize the information of adjacent items when decoding words.

[0086] The present invention verifies the proposed adjacent information enhancement method on the basis of the conversational recommendation system C 2 -CRS. The test set is the dataset ReDial (Li R, Kahou S E, Schulz H, etc. Towards deep conversational recommendations[J]. Advances in Neural Information Processing Systems, 2018, 2018 - December(NeurIPS): 9725–9735.), which is widely used in the field of conversational recommendation. Introduce the performance of the method of the present invention in the item recommendation task and the response generation task.

[0087] In the item recommendation task, the evaluation metrics are Recall@K, MRR@K, and NDCG@K, which are commonly used to measure the recommendation list. In addition, the present invention also selects 8 commonly used baseline methods in the conversational recommendation system for comparison. It can be found from the results in Table 1 that the present invention enables the conversational recommendation system C 2 -CRS to have a significant improvement in all metrics and outperforms all baseline methods.

[0088] Table 1

[0089]

[0090] In the response generation task, the present invention adopts the automatic evaluation metric Distinct@N, the manual evaluation metrics fluency and informativeness. It can be found from the results in Table 2 that the present invention enables C 2- The CRS has achieved significant improvements in all metrics. Distinct@N measures the diversity of the generated responses, while informativeness aims to evaluate how much meaningful information is contained in the generated responses. The Transformer method is not a conversational recommendation method. It is a classical response generation baseline method that does not consider item information. Therefore, the Transformer method performs best in terms of fluency and worst in terms of informativeness. By using the method of the present invention, good performance can be achieved in both fluency and informativeness.

[0091] Table 2

[0092]

[0093] The method of the present invention can be directly applied to an open-domain chatbot system and is a core module of a chatbot.

[0094] First, the central control module hands over the input and control to the item recommendation module, which generates a response containing recommended items for the central control module to complete a conversational recommendation task.

[0095] In terms of the deployment method, this technology can be independently deployed as a computing node on cloud computing platforms such as Alibaba Cloud or Meituan Cloud, and the communication with other modules can be carried out by binding IP addresses and port numbers.

[0096] In the specific implementation of this technology, because deep learning-related technologies are used, corresponding deep learning frameworks are required: the relevant experiments of the present invention are implemented based on the open-source framework Pytorch. It can also be replaced with other frameworks, such as the equally open-source tensorflow, or PadlePadle used within the enterprise.

[0097] The above examples of the present invention are only to illustrate in detail the calculation model and calculation process of the present invention, rather than to limit the implementation manner of the present invention. For those of ordinary skill in the art, other different forms of changes or variations can be made based on the above description. It is impossible to list all the implementation manners here. Any obvious changes or variations derived from the technical solution of the present invention are still within the protection scope of the present invention.

Claims

1. A method for enhancing a conversational recommendation system using adjacent information, characterized in that, the method specifically includes the following steps: Step 1: Construct pseudo-label data using adjacent items of the target item to be recommended in the conversation, and then construct a training set for the conversational recommendation system based on the pseudo-label data. After training, a trained conversational recommendation system is obtained; The specific process of Step 1 is: Step 1: Obtain a conversation dataset; Step 2: For any conversation in the dataset, construct a training sample using this conversation Among them, represents the representation vector of the context of this conversation obtained by the conversational recommendation system, represents the representation vector of the target item recommended for this conversation, represents the representation vector of the response generated for this conversation; Using the representation vector of the target item Retrieve n items adjacent to the target item from the candidate item set Y to be recommended, and use the representation vectors of the retrieved adjacent items to replace the training samples in The replaced training samples are used as pseudo-label training samples; Similarly, each conversation in the dataset is processed separately. Denote the set composed of training samples constructed based on each conversation in the dataset as D, and denote the set of all obtained pseudo-labeled training samples as where represents the set of adjacent items retrieved according to the representation vector of a target item, represents the set the number of adjacent items contained in, that is |D| represents the number of conversations contained in the dataset, and i' represents the i'-th pseudo-labeled training sample in D p the i'-th pseudo-labeled training sample in represents the representation vector of the conversation context corresponding to the i'-th pseudo-labeled training sample, represents the representation vector of the recommended target item in the i'-th pseudo-labeled training sample, represents the representation vector for generating a response to the conversation corresponding to the i'-th pseudo-labeled training sample; Step 3: Take the union of D and D p as D ∪ D p as the training set for this iteration, and use the training set to train the conversational recommendation system Step 4. Repeat the process of Step 2 to Step 3 until the iteration stops when the set maximum number of iterations is reached, and a trained conversational recommendation system is obtained. Step 2: Input the current conversation into the trained conversational recommendation system to obtain an item recommendation result and a response generation result for the current conversation.

2. A method for enhancing a conversational recommendation system using adjacent information according to claim 1, characterized in that, The representation vector of the target item Retrieve n items adjacent to the target item from the candidate item set Y to be recommended. The specific process is as follows: Calculate the cosine similarity between the representation vectors of each item in the candidate item set Y to be recommended and respectively, and take the n items with the largest cosine similarity to as the n adjacent items retrieved: Among them, the function represents retrieving n adjacent items from the set Y of candidate items to be recommended.

3. A method for enhancing a conversational recommendation system using adjacent information according to claim 2, characterized in that, The specific process of Step 3 is: Step 1): For any training sample in the union set D∪D p retrieve, from the candidate dialogue representation vector set X, the dialogue representation vector adjacent to the context representation vector of this training sample calculate the parameter b′ of each candidate item and the vector corresponding to this training sample based on the retrieved adjacent dialogue representation vectors According to the vector and the parameter b', obtain the item recommendation result and the response generation result for the context representation vector ; Similarly, each training sample in the union set D∪D p is processed; Step 2): Calculate the loss of the item recommendation task based on the item recommendation result and the response generation result obtained in Step 1), and then update the parameters of the conversational recommendation system according to the loss. ; Step 3): Set the pseudo-label training sample set to be an empty set, that is, set Then return to Step 2.

4. A method for enhancing a conversational recommendation system using adjacent information according to claim 3, characterized in that, The calculation process of the vector x' is: Given a set of candidate dialogue context representation vectors \(X\), calculate the cosine similarity between each vector in \(X\) and the vector . Take the \(m\) dialogues corresponding to the largest \(m\) cosine similarities as the adjacent dialogues of the vector ; Among them, represents a set of adjacent conversations, and φ(x, X, m) represents retrieving m conversations adjacent to the vector from the set X; Among them, v i is the intermediate vector of the i-th adjacent conversation, w i is the weight of the i-th adjacent conversation, represents the set the number of adjacent conversations in, R represents the set of real numbers, and d is the dimension of.

5. A method for enhancing a conversational recommendation system using adjacent information according to claim 4, characterized in that, The said v i According to the representation vector of the i-th adjacent conversation and the representation vector of the recommended item of the i-th adjacent conversation Calculate: where MLP is a multi-layer perceptron.

6. A method for enhancing a conversational recommendation system using adjacent information according to claim 5, characterized in that, The said w i According to the representation vector of the i-th adjacent conversation and calculate: where Linear is a linear layer and Sigmoid is an activation function.

7. A method for enhancing a conversational recommendation system using adjacent information according to claim 6, characterized in that, The MLP consists of two linear layers.

8. A method for enhancing a conversational recommendation system using adjacent information according to claim 7, characterized in that, The parameter b ′ is calculated as follows: Among them, b and b s are intermediate parameters.

9. A method for enhancing a conversational recommendation system using adjacent information according to claim 8, characterized in that, The intermediate parameters b and b s are calculated as follows: For candidate item e: Among them, τ k represents the number of times the candidate item e appears in the k-th conversation, and w k represents the weight of the k-th conversation, and the weight of the current conversation is 1; where, T k,j represents the number of times the j-th item appears in the k-th conversation, E is the item set, |E| is the number of items in the item set, s j is the cosine similarity between the representation vector of the j-th item in the item set and the representation vector of the candidate item e; Take b and b s Do and obtain the parameter b corresponding to the candidate item e ′ ; Similarly, the parameter b corresponding to each item in the item set E is obtained ′ .

10. A method for enhancing a conversational recommendation system using adjacent information according to claim 9, characterized in that, The specific process of Step 2 is: Step ①: Calculate the probability of being recommended for each candidate item respectively: Among them, represents the context representation vector of the current conversation, and the superscript T represents transpose. represents the representation vector of the candidate item. represents the vector is the probability that the item corresponding to is recommended. The candidate item with the highest recommended probability is used as the item recommendation result; Step ②: After obtaining the recommended item, retrieve the adjacent items of the recommended item and generate a response based on the retrieved adjacent items. Among them, the Decoder is a decoder that selects Q b from the retrieved adjacent items ′ The item with the largest parameter value, and fuses the representation vectors of the selected items to obtain a fusion result are the 1st, …, (q - 1)th word vectors in the generated response, indicating the probability of being the qth word vector in the generated response; The maximum probability The corresponding word vector is used as the q-th word vector for generating the response.

Images